Automatic anomaly detection in fuel grab load trace data using a knowledge-based system vs. multiple deep autoencoders

Research output: Contribution to conferencePaper

Abstract

Of the seven Advanced Gas-cooled Reactor nuclear power stations in the UK, the majority are approaching their planned closure date. As the graphite core of these type of reactors cannot be repaired or replaced, this is one of the main life-limiting factors. The refuelling of a nuclear power station is an ongoing process and refuelling of the reactor occurs typically every 6 to 8 weeks. During this process, data relating to the weight of the fuel assembly is recorded: this data is called fuel grab load trace data and the major contributing factor to this are the frictional forces, with a magnitude related to the channel bore diameter. Through an understanding of this data, it is possible to manually interpret whether there are any defects in the individual brick layers that make up the graphite core but doing so requires significant expertise, experience and understanding.
In this paper, we present a knowledge-based system to automatically detect defects in individual brick layers in the fuel grab load trace data. This is accomplished using a set of rules defined by specialist engineers. Secondly, by splitting up the trace into overlapping regions, the use of multiple deep autoencoders is explored to produce a generative model for a normal response. Using this model, it is possible to detect responses that do not generalise and identify anomalies such as defects in the individual brick layers. Finally, the two approaches are compared, and conclusions are drawn about the applications of these techniques into industry.

Conference

Conference11th Nuclear Plant Instrumentation, Control and Human-Machine Interface Technologies (NPIC&HMIT) 2019
CountryUnited States
CityOrlando
Period9/02/1914/02/19

Fingerprint

Knowledge based systems
Brick
Defects
Nuclear power plants
Graphite
Gas cooled reactors
Engineers
Industry

Keywords

  • fuel grab load trace
  • machine learning
  • autoencoders
  • condition monitoring
  • advanced gas-cooled reactor (AGR)

Cite this

Young, A. T., Aylward, W., Murray, P., West, G. M., & McArthur, S. D. J. (2019). Automatic anomaly detection in fuel grab load trace data using a knowledge-based system vs. multiple deep autoencoders. Paper presented at 11th Nuclear Plant Instrumentation, Control and Human-Machine Interface Technologies (NPIC&HMIT) 2019, Orlando, United States.
Young, A. T. ; Aylward, W. ; Murray, P. ; West, G. M. ; McArthur, S. D. J. / Automatic anomaly detection in fuel grab load trace data using a knowledge-based system vs. multiple deep autoencoders. Paper presented at 11th Nuclear Plant Instrumentation, Control and Human-Machine Interface Technologies (NPIC&HMIT) 2019, Orlando, United States.10 p.
@conference{9aba4b0b23c546c2acfd425a7b43ccdb,
title = "Automatic anomaly detection in fuel grab load trace data using a knowledge-based system vs. multiple deep autoencoders",
abstract = "Of the seven Advanced Gas-cooled Reactor nuclear power stations in the UK, the majority are approaching their planned closure date. As the graphite core of these type of reactors cannot be repaired or replaced, this is one of the main life-limiting factors. The refuelling of a nuclear power station is an ongoing process and refuelling of the reactor occurs typically every 6 to 8 weeks. During this process, data relating to the weight of the fuel assembly is recorded: this data is called fuel grab load trace data and the major contributing factor to this are the frictional forces, with a magnitude related to the channel bore diameter. Through an understanding of this data, it is possible to manually interpret whether there are any defects in the individual brick layers that make up the graphite core but doing so requires significant expertise, experience and understanding. In this paper, we present a knowledge-based system to automatically detect defects in individual brick layers in the fuel grab load trace data. This is accomplished using a set of rules defined by specialist engineers. Secondly, by splitting up the trace into overlapping regions, the use of multiple deep autoencoders is explored to produce a generative model for a normal response. Using this model, it is possible to detect responses that do not generalise and identify anomalies such as defects in the individual brick layers. Finally, the two approaches are compared, and conclusions are drawn about the applications of these techniques into industry.",
keywords = "fuel grab load trace, machine learning, autoencoders, condition monitoring, advanced gas-cooled reactor (AGR)",
author = "Young, {A. T.} and W. Aylward and P. Murray and West, {G. M.} and McArthur, {S. D. J.}",
year = "2019",
month = "2",
day = "12",
language = "English",
note = "11th Nuclear Plant Instrumentation, Control and Human-Machine Interface Technologies (NPIC&HMIT) 2019 ; Conference date: 09-02-2019 Through 14-02-2019",

}

Young, AT, Aylward, W, Murray, P, West, GM & McArthur, SDJ 2019, 'Automatic anomaly detection in fuel grab load trace data using a knowledge-based system vs. multiple deep autoencoders' Paper presented at 11th Nuclear Plant Instrumentation, Control and Human-Machine Interface Technologies (NPIC&HMIT) 2019, Orlando, United States, 9/02/19 - 14/02/19, .

Automatic anomaly detection in fuel grab load trace data using a knowledge-based system vs. multiple deep autoencoders. / Young, A. T.; Aylward, W.; Murray, P.; West, G. M.; McArthur, S. D. J.

2019. Paper presented at 11th Nuclear Plant Instrumentation, Control and Human-Machine Interface Technologies (NPIC&HMIT) 2019, Orlando, United States.

Research output: Contribution to conferencePaper

TY - CONF

T1 - Automatic anomaly detection in fuel grab load trace data using a knowledge-based system vs. multiple deep autoencoders

AU - Young, A. T.

AU - Aylward, W.

AU - Murray, P.

AU - West, G. M.

AU - McArthur, S. D. J.

PY - 2019/2/12

Y1 - 2019/2/12

N2 - Of the seven Advanced Gas-cooled Reactor nuclear power stations in the UK, the majority are approaching their planned closure date. As the graphite core of these type of reactors cannot be repaired or replaced, this is one of the main life-limiting factors. The refuelling of a nuclear power station is an ongoing process and refuelling of the reactor occurs typically every 6 to 8 weeks. During this process, data relating to the weight of the fuel assembly is recorded: this data is called fuel grab load trace data and the major contributing factor to this are the frictional forces, with a magnitude related to the channel bore diameter. Through an understanding of this data, it is possible to manually interpret whether there are any defects in the individual brick layers that make up the graphite core but doing so requires significant expertise, experience and understanding. In this paper, we present a knowledge-based system to automatically detect defects in individual brick layers in the fuel grab load trace data. This is accomplished using a set of rules defined by specialist engineers. Secondly, by splitting up the trace into overlapping regions, the use of multiple deep autoencoders is explored to produce a generative model for a normal response. Using this model, it is possible to detect responses that do not generalise and identify anomalies such as defects in the individual brick layers. Finally, the two approaches are compared, and conclusions are drawn about the applications of these techniques into industry.

AB - Of the seven Advanced Gas-cooled Reactor nuclear power stations in the UK, the majority are approaching their planned closure date. As the graphite core of these type of reactors cannot be repaired or replaced, this is one of the main life-limiting factors. The refuelling of a nuclear power station is an ongoing process and refuelling of the reactor occurs typically every 6 to 8 weeks. During this process, data relating to the weight of the fuel assembly is recorded: this data is called fuel grab load trace data and the major contributing factor to this are the frictional forces, with a magnitude related to the channel bore diameter. Through an understanding of this data, it is possible to manually interpret whether there are any defects in the individual brick layers that make up the graphite core but doing so requires significant expertise, experience and understanding. In this paper, we present a knowledge-based system to automatically detect defects in individual brick layers in the fuel grab load trace data. This is accomplished using a set of rules defined by specialist engineers. Secondly, by splitting up the trace into overlapping regions, the use of multiple deep autoencoders is explored to produce a generative model for a normal response. Using this model, it is possible to detect responses that do not generalise and identify anomalies such as defects in the individual brick layers. Finally, the two approaches are compared, and conclusions are drawn about the applications of these techniques into industry.

KW - fuel grab load trace

KW - machine learning

KW - autoencoders

KW - condition monitoring

KW - advanced gas-cooled reactor (AGR)

UR - http://npic-hmit.ans.org/

M3 - Paper

ER -

Young AT, Aylward W, Murray P, West GM, McArthur SDJ. Automatic anomaly detection in fuel grab load trace data using a knowledge-based system vs. multiple deep autoencoders. 2019. Paper presented at 11th Nuclear Plant Instrumentation, Control and Human-Machine Interface Technologies (NPIC&HMIT) 2019, Orlando, United States.